Africa’s ecosystems exhibit a tradeoff between resistance and stability following disturbances

Environmental disturbances may prevent ecosystems from consistently performing their critical ecological functions. Two important properties of ecosystems are their resistance and stability, which respectively reflect their capacities to withstand and recover from disturbance events (e.g. droughts, wildfires, pests, etc). Theory suggests that resistant and stable ecosystems possess opposing characteristics, but this has seldom been established across diverse ecosystem attributes or broad spatial scales. Here, we compare the resistance and stability of >1000 protected area ecosystems in Africa to disturbance-induced losses in primary productivity from 2000 to 2019. We quantitatively evaluated each ecosystem such that following disturbances, an ecosystem is more resistant if it experiences lower-magnitude losses in productivity, and more stable if it returns more rapidly to pre-disturbance productivity levels. To compare the characteristics of resistant versus stable ecosystems, we optimized random forest models that use ecosystem attributes (representing their climatic and environmental conditions, plant and faunal biodiversity, and exposure to human impacts) to predict their resistance and, separately, stability values. We visualized each attribute’s relationship with resistance and stability after accounting for all other attributes in the model framework. Ecosystems that are more resistant to disturbances are less stable, and vice versa. The ecosystem attributes with the most predictive power in our models all exhibit contrasting relationships with resistance versus stability. Notably, highly resistant ecosystems are generally more arid and exhibit high habitat heterogeneity and mammalian biodiversity, while highly stable ecosystems are the opposite. We discuss the underlying mechanisms through which these attributes engender resistance or, conversely, stability. Our findings suggest that resistance and stability are fundamentally opposing phenomena. A balance between the two must be struck if ecosystems are to maintain their identity, structure, and function in the face of environmental change.


Introduction
A lasting ecosystem is one that can endure dramatic environmental change [1]. It can effectively withstand and recover from disturbances, consistently maintaining its identity, structure, and function over time [2][3][4][5][6]. The capacity of ecosystems to overcome disturbance today is increasingly under threat, as landscapes are impacted by rapid biodiversity loss, climate change, altered fire regimes, and other major human influences [1,7,8]. If we are to ensure that ecosystems continue to perform their essential ecological functions, we must understand how they are affected by and respond to disturbance events.
Ecosystems possess multiple properties that characterize their responses to disturbance [9][10][11][12], and among the most recognized are resistance and stability [13]. Resistance refers to the capacity of an ecosystem to withstand disturbances [2,4,6,14,15]. When exposed to a disturbance event, like a fire or an invasive species, a resistant ecosystem experiences minimal losses to its identity, structure, and function. Stability (which many studies alternatively call 'resilience' or 'engineering resilience') refers to the rate at which an ecosystem can recover to its baseline level of function following a disturbance, where a more stable ecosystem recovers more quickly [2, 4-6, 13, 14, 16-18]. Without both resistance and stability, ecosystems may transition to depauperate ecological states with depleted function [19]. For instance, a forest ecosystem exposed to a wildfire could experience major declines in primary productivity if minimally resistant, as well as slow and potentially incomplete regrowth if minimally stable. In either case, the forest could transition to a barren, low-biodiversity state that is less capable of performing critical processes [17,20] like carbon sequestration [21].
Although both resistance and stability are necessary to maintain ecosystem function, theory suggests that resistant ecosystems possess characteristics that oppose those of stable ecosystems [22][23][24]. Among the most well-established of these characteristics is an ecosystem's degree of connectivity between individuals and/or habitat areas. With greater connectivity comes reduced resistance, as a disturbance that impacts some individuals or areas can more easily spread to others [25][26][27][28][29]. However, greater connectivity also fosters stability, as organisms can then more easily traverse space and assist in the re-population of degraded regions [26,28]. The result is that ecosystem characteristics, like connectivity, cannot necessarily bolster both resistance and stability, requiring a balance between the two to maintain ecosystems.
While this resistance-stability tradeoff has been explored, it has seldom been established across a variety of ecosystem characteristics or on broad spatial scales. Studies of resistance and stability often focus on a subset of attributes and forces that describe ecosystems, like plant biodiversity and climate [11,22,30]. Few studies, if any, simultaneously consider the effects of climate, environmental conditions, human impacts, and biodiversity (across multiple trophic levels and dimensions). Further, site-level and fieldbased assessments of stability are more prevalent in the literature than broad-scale studies that span multiple biomes [31]. Thus, we have not yet established if the resistance-stability tradeoff is a widespread phenomenon, or if it is specific to only certain ecosystem attributes and localities.
Evaluating this tradeoff is particularly important for protected area ecosystems in Africa. Protected areas form the foundation of in situ biodiversity conservation [32,33], and in Africa they house some of the most unique and biodiverse faunal assemblages on Earth [34,35]. While protected areas globally are often placed in regions with low agricultural value [36], many in Africa were positioned by colonial powers to conserve species considered charismatic or aesthetic [32,[37][38][39]. This species-level focus, while valuable, is not as effective as ecosystem-level approaches in detecting broad ecological patterns that influence the persistence of many taxa [40]. Ecosystem-level initiatives, like programs for volunteers to monitor savannah function [41], are essential if Africa's protected areas are to conserve biodiversity moving forward. Supporting these initiatives requires that we advance our understanding of ecosystem processes, including the resistance-stability tradeoff. This is critical now, at a time when Africa's wildlife populations are isolated [42] and in decline [43], and its ecosystems are threatened by pests, diseases [44], droughts [45], and other major disturbances.
Here, we evaluate if resistance and stability oppose each other across >1000 protected area ecosystems in Africa. Using enhanced vegetation index (EVI) data, we designate ecosystems as more resistant if, from 2000 to 2019, they experienced lowermagnitude losses in primary productivity following disturbance events, and as more stable if they subsequently recovered more quickly to baseline productivity levels. We assess spatial patterns in resistance and stability, as well as their correlation. We then train models to predict the resistance and stability values of ecosystems, and to determine how a suite of ecosystem attributes relates with each of resistance and stability. If the attributes' relationships with one are opposite to their relationships with the other, then the characteristics of resistant ecosystems fundamentally counter those of stable ones. We hypothesize that such a tradeoff broadly exists [25][26][27][28].

Selection of protected area ecosystems
To select protected areas for our study, we obtained and processed polygons representing the geographic extents of protected areas in Africa from the World Database on Protected Areas [46]. We only considered those that became recognized before the year 2000, as the temporal range used to measure ecosystem stability begins in 2000. Of the remaining protected areas, we only retained those large enough to encompass a 5 × 5 km raster pixel at its centroid. This allowed us to examine ecosystems at relatively broad (5 × 5 km) and small (1 × 1 km) scales [47], and in regions more likely to be far from protected area boundaries. A threshold larger than 5 × 5 km entailed the removal of too many protected areas for a meaningfully large sample size. We compiled a finalized list of 1384 protected areas spanning Africa.
Our study analyzed ecosystems occurring within these 1384 protected areas. For each protected area, we selected the 5 × 5 km (or, separately, 1 × 1 km) raster pixel occurring at its centroid and designated that pixel as an ecosystem. Selecting a single ecosystem per protected area circumvents the spatial autocorrelation that emerges between adjacent regions. To account for spatial scale, we repeated all methods that follow separately for both 1 × 1 km and 5 × 5 km raster pixels.

Collection of ecosystem attribute data
We collected data representing attributes of the selected ecosystems that are hypothesized drivers of ecosystem resistance and stability. We chose attributes that comprehensively span a variety of biotic and abiotic conditions. A detailed description of how these attributes were obtained and quantified, and why each was chosen, is in appendix S1.1 and table S1 of the supplementary material.
In addition to these attributes, we used EVI measurements to quantify each ecosystem's resistance and stability. EVI captures an ecosystem's level of net primary productivity, a proxy for its degree of functioning and a powerful tool for broad-scale assessments of resistance and stability [4]. The NASA MOD13A2 product provides global raster layers that contain the EVI of 1 × 1 km terrestrial pixels at 16 day intervals from 2000 to 2019 [48]. We obtained all such layers, and we generated 5 × 5 km versions of them by aggregating the 1 × 1 km layers by their mean pixel values. For each resolution, we extracted all EVI values across the 20 years at each protected area ecosystem centroid, resulting in a time series of values per ecosystem. The following methods use these time series to quantify resistance and stability, based on White et al [4]. A detailed justification of these methods is in appendix S2.
Time series of EVI values contain seasonal cycles that could interfere with resistance and stability signals [49] and must be removed. We accounted for these cycles within each ecosystem's EVI time series, at each resolution, by converting its EVI values into unitless anomalies [4]. We did so for each EVI value as follows: (1) identify the 16 day interval that the value represents (e.g. 1-16 January); (2) calculate the mean and standard deviation of all 20 values from 2000 to 2019 pertaining to that interval; and (3) subtract the mean from the EVI value, and then divide the result by the standard deviation. An EVI anomaly of 0 represents the baseline level of ecosystem function, i.e. the expected level of function at a point in time given the ecosystem's seasonal cycles. The conversion of EVI values to anomalies is required for valid comparisons of ecosystems. Without it, we would falsely conclude that an ecosystem with naturally highermagnitude seasonal fluctuations in EVI is less resistant than one with lower-magnitude fluctuations (e.g. a forest vs. desert ecosystem; see appendix S2.1).
Because resistance and stability represent an ecosystem's response to disturbances, we identified disturbance events within each ecosystem's time series of EVI anomalies at each resolution. We defined a disturbance event as one associated with an EVI anomaly value <−1. This represents a loss in ecosystem function that is greater than one standard deviation away from baseline function (i.e. a one-sigma event [4]). The threshold of −1 provides a sufficient sample of distinct disturbance events in each ecosystem (minimum = 5 and mean = 13.5 across all ecosystems and resolutions), allowing resistance and stability to be assessed in all ecosystems. Further justification of the EVI threshold is in appendix S2.2. Increasing the magnitude of the threshold from −1 to −1.5 or −2 was not feasible, as it captured too few disturbances for a signal to be found in our downstream regression models. However, decreasing the threshold to −0.5 led to similar results as those of −1, indicating that our results are minimally sensitive to the chosen threshold (see appendix S2.3).
We calculated each ecosystem's resistance and stability from its disturbance events. First, we identified each ecosystem's one-sigma events that refer to distinct disturbances (and not continuations of others), using the following strategy [4]: (1) designate the ecosystem's highest-magnitude EVI anomaly that is <−1 as a distinct one-sigma event; (2) remove that anomaly and the 20 anomaly values before and after it. Twenty refers to the number of 16 day intervals required for 97.5% of one-sigma events to return to an anomaly of ⩾0 across all ecosystems [4]; (3) return to step 1 for the next highest-magnitude event remaining. Then, we quantified each ecosystem's resistance as follows: Here, n is the number of one-sigma events that occur, and E is the EVI anomaly value of each event. Each ecosystem's resistance is then the mean of the reciprocals of all its one-sigma event magnitudes [4]. A higher value denotes a more resistant ecosystem with lower-magnitude losses in function following disturbance. Finally, we quantified each ecosystem's stability (called 'engineering resilience' in White et al [4]) as follows: Here, n and E are the same as in the resistance equation, and T is the time taken after each one-sigma event to recover to baseline function. We calculated T for each one-sigma event as the number of subsequent 16 day intervals it takes for a 48 day moving window of EVI anomalies to return to a mean ⩾0. The moving window accounts for potential noise in the anomalies [4]. The stability of each ecosystem is finally defined as its mean recovery rate across all its one-sigma events, where a higher rate denotes greater stability. These formulations align with prior literature that describes resistance as the degree to which a system's state is displaced following a disturbance [2,50], and stability as a system's subsequent rate of recovery [5,18,51]. They also divulge real patterns and not statistical artefacts (see appendix S2.4).

Data exploration and pre-processing
We first analyzed the resistance and stability values of all 1384 ecosystems. At each spatial resolution, we calculated and visualized the Spearman's rank correlation coefficient (Spearman's ρ) between resistance and stability, to assess the strength and direction of their relationship. Further, we mapped the ecosystems' resistance and stability values across Africa to assess their distribution across space.
We pre-processed our two datasets compiled above, one representing ecosystems at 1 × 1 km resolution and the other ecosystems at 5 × 5 km resolution. We removed the few ecosystems that possessed missing data, most commonly with respect to their plant biodiversity (table S1, green shading). This resulted in a dataset of 1145 ecosystems at 1 × 1 km resolution and 1120 at 5 × 5 km resolution. Both datasets were analyzed identically moving forward. We assessed if multicollinearity exists between all attributes in table S1 via variance inflation factor (VIF) analysis. VIF assesses, per attribute, the degree to which its variance is explained by the other attributes present. For a given attribute, a VIF value of 10 or higher, or even as low as 4, is often considered a sign of severe multicollinearity and grounds for the attribute's removal from analysis, although these thresholds may be too stringent [52]. All attributes in both 1 × 1 km and 5 × 5 km-resolution datasets incurred VIF values ⩽∼3, indicating that severe multicollinearity does not exist in our data. Similarly, we determined that no interaction terms between attributes are necessary for our regression models (see appendix S1.2). Next, using the 'spatialRF' package in R [53,54], we calculated the pair-wise distances between ecosystems to use in evaluating our regression models for spatial autocorrelation. We also Box-Cox transformed all continuous attributes in table S1, which normalizes their distributions and mitigates the impacts of outlier data points on model performance. Finally, we split each of our two datasets into model training and testing subsets, where each training subset comprised 80% of ecosystems and each testing subset 20%. We used the 'blockCV' package in R [55] to perform the split in a way that limits the spatial non-independence of the two subsets (see appendix S1.3) [55,56].

Modelling ecosystem resistance and stability
We optimized four regression models, one for each of the 1 × 1 km and 5 × 5 km-resolution datasets, and one for each of ecosystem resistance and stability as the response variable. Each model used the attributes in table S1 to predict its response, and then to assess the marginal relationship between each attribute and the response. To make all models comparable, we repeated the following methods identically for each.
With the 'caret' package in R [57], we trained each regression model by applying a regularized random forest (RRF) to its training data subset. RRF is a type of random forest model that emphasizes the use of the most predictive attributes [58], enhancing model robustness. We used a five-fold crossvalidation scheme in model training to identify the hyperparameters of the RRF that led to the best predictions of the response. These parameters included the number of random attributes to consider at each node of the random forest's decision trees, and the coefficient of regularization, where a lower coefficient value more heavily penalizes using attributes with minimal predictive power [58]. All four models were best fit with few attributes per node (∼4) and a low coefficient value (∼0.1), so for consistency we used these same hyperparameters in each model.
We evaluated, per model, whether the residuals of its predictions on the response during training were subject to spatial autocorrelation. We did so with the Moran's I metric. Moran's I uses the priorly constructed distance matrix between ecosystems as input and outputs a value between −1 and 1, where 0 denotes no autocorrelation. All four models produced a Moran's I value of <0.015, denoting that their predictions are subject to minimal spatial autocorrelation. Therefore, no corrections were necessary to ensure that the models treat ecosystems properly as independent data points across space.
In training each model, we assessed the strength and direction of the relationship between each attribute and the response variable. We assessed strength by calculating each attribute's importance in predicting the response. Per attribute, we used the 'vip' package in R [59] to (1) permute its values across all ecosystems in the training dataset 100 times; and (2) for each permutation, to calculate the decline in the R-squared of the model relative to when the values were not permuted. An attribute associated with a greater median decline across all 100 permutations is of greater importance. After ranking all attributes by their importance, we calculated the mean rank of each group of attributes in table S1 to compare groups. We assessed the direction of relationships using partial dependence plots, which depict how the values of the response variable change across the values of an attribute, after accounting for the effects of all other attributes in the trained RRF model. We used the 'pdp' package in R [60] to produce the data for these plots, i.e. plots of the expected values of the response variable versus the values of each numeric attribute. We fit LOESS regression curves to all plots to visualize their trends.
The reliability of the trained models depends on their predictive performance. We evaluated the performance of each trained model using each's corresponding testing data subset. First, we used the trained model to predict the values of the response variable in the testing subset. We then assessed the degree to which each model's predictions matched corresponding actual values. We calculated the predictions' normalized root mean squared error (RMSE), equal to the predictions' RMSE divided by the mean of the actual values. A model with a normalized RMSE of ⩽0.2 is considered to have good precision [61]. We plotted the actual versus predicted values of each model's response variable and assessed the slope, intercept, and R-squared of the plot's regression line. We performed all analyses in R 3.6.1 [62].

Results
The resistance and stability of Africa's protected area ecosystems are inversely related. Regions across the continent contain ecosystems that tend to be high in resistance (figures 1(a) and S1(a)) and low in stability (figures 1(b) and S1(b)), like southern mainland Africa, or vice versa, as along the east coast of Madagascar (figures 1 and S1). Further, resistance and stability values are negatively correlated across all ecosystems (ρ = −0.56 at 5 × 5 km and −0.60 at 1 × 1 km resolution; figures 2 and S2). Ecosystems less capable of withstanding disturbances, in that they experience higher-magnitude losses in primary productivity following disturbance events (low resistance), can subsequently recover more quickly to baseline function (high stability). Meanwhile, ecosystems more capable of withstanding disturbances (high resistance) recover more slowly (low stability).
The regression models built at 5 × 5 km resolution exhibited greater R-squared values than those built at 1 × 1 km resolution, although all models exhibited good precision (normalized RMSE ⩽ ∼0.2). The model trained to predict resistance at 5 × 5 km resolution produced an R-squared of 43.4% and normalized RMSE of 0.091 in its predictions of the resistance values in the testing data ( figure  S3(a)). At 1 × 1 km resolution, the corresponding R-squared was 38.1% and normalized RMSE 0.092 ( figure S3(b)). Meanwhile, the models for ecosystem stability exhibited an R-squared of 40.1% and normalized RMSE of 0.230 at 5 × 5 km resolution (figure S3(c)), and 31.8% and 0.215 at 1 × 1 km resolution ( figure S3(d)).
All attributes in table S1 were of importance in predicting ecosystem resistance and stability at both 5 × 5 km (figure 3) and 1 × 1 km ( figure  S4) resolutions, although levels of importance varied. Each attribute raised the R-squared value of each trained model significantly when not permuted versus permuted, as none of the boxplots or outliers in figure 3 or figure S4 overlap with an importance value ⩽0. However, the most important attributes across all models were those representing climate (figures 3 and S4, blue plots), plant and mammal biodiversity (green and purple plots, respectively), and environmental conditions (gold), particularly landscape heterogeneity. The attributes representing human impacts (orange), as well as protected area size and shape (gray), were comparatively less important (table S2).
Partial dependence plots from our models display inverse patterns with respect to how each numeric attribute in table S1 relates to ecosystem resistance versus stability. Attributes that relate positively with resistance tend to relate negatively with stability, and vice versa, at both 5 × 5 km and 1 × 1 km resolutions (figures 4 and S5-9). This is particularly true among the more important attributes (figures 3 and S4), including those we use as examples for further discussion (figures 4 and S9). For instance, ecosystems subject to high annual precipitation are less resistant, in that they experience higher-magnitude losses in function following disturbance; however, they are also more stable, in that they recover more quickly to baseline function (figures 4 and S9).

Discussion
A strong tradeoff exists between the resistance and stability of Africa's protected area ecosystems. After exposure to disturbance events, resistant ecosystems that elude major declines in function (i.e. primary productivity) are generally less stable, as they are slower to recover to pre-disturbance levels of function, and vice versa (figures 2 and S2). Of the ecosystem attributes analyzed (table S1), the majority substantiate that the characteristics of resistant ecosystems oppose those of stable ecosystems. That is, the relationships that each attribute shares with resistance versus stability are largely mirror images of each other (figures 4 and S5-9). These results are based on reliable regression model performance ( figure S3) and, contrary to the sensitivity of ecosystem processes to spatial scale [47,63], are highly similar at both 1 × 1 km and 5 × 5 km resolutions.
Consequently, we find that resistance and stability appear to be fundamentally opposing properties of ecosystems across multiple spatial resolutions, diverse ecosystems, and numerous biotic and abiotic forces. An ecosystem whose characteristics maximize its resistance tend to minimize its stability, and vice versa, making it unlikely that any combination of characteristics can simultaneously maximize both. This necessitates a balance between the two, as both resistance and stability are essential if Africa's protected area ecosystems are to avoid transitioning to depauperate ecological states with depleted function [17,19,20].
The attributes of Africa's ecosystems (table S1) highlight the mechanisms underlying this resistancestability tradeoff by demonstrating the circumstances in which an ecosystem is maximally resistant or, conversely, stable. For simplicity, we focus our discussion on four attributes, one from each of the groups of relatively high importance in our regression models   56), meaning that ecosystems less capable of withstanding disturbances (low resistance) recover more quickly following disturbances (high stability), and vice versa.
( figure 3; table S2). A maximally resistant ecosystem is one that occurs in a region of relatively limited plant productivity (even after accounting for the biome in which it occurs), with minimal annual precipitation (figures 4(a) and S9(a)) [64,65] and vegetation that possesses low leaf dry matter content (figures 4(b) and S9(b)) [66]. The ecosystem's low-productivity vegetation is less reliant on water availability for survival, making it more capable of withstanding resource-limiting disturbances like droughts [67,68]. Further, the ecosystem occurs in a region of high habitat heterogeneity (figures 4(c) and S9(c)). Traversing space requires movement through diverse habitat types, each with unique abiotic conditions [69]. This means that disturbances like diseases [44] are a limited threat because of their reduced capacity to spread. The ecosystem's heterogeneous surroundings also engender modular species interaction networks, reducing the risk of disturbance-induced extinction cascades [70]. Finally, this maximally resistant ecosystem comprises a mammalian community with high functional diversity (figures 4(d) and S9(d)). Its mammals perform critical functions like seed dispersal and habitat structuring [71,72], and because of their high functional diversity, exhibit more varied responses to disturbance [73][74][75]. Thus, during a disturbance, at minimum a subset of mammalian species persists and continues contributing to the ecosystem's function. Many ecosystems, like those occurring in Namibia's Naute Recreation Resort and Torra Conservancy, as well as other protected areas throughout southern Africa (figures 1(a) and S1(a)), possess these resistance-inducing characteristics.
A maximally stable ecosystem, conversely, is one that possesses characteristics that are largely the opposite. Unlike the resistant ecosystem, it is located in a region of high annual precipitation (figures 4(a) and S9(a)), where precipitation promotes the rapid post-disturbance regrowth of plants [49,76,77] and the emergence of new ones from seed banks [11]. It possesses vegetation with high leaf dry matter content (figures 4(b) and S9(b)), indicative of a greater capacity to conserve nutrients and to store resources from prior favorable conditions [78,79]. Consequently, its vegetation is more equipped to efficiently regenerate following disturbance events. This maximally stable ecosystem also occurs in a region where habitat heterogeneity is low (figures 4(c) and S9(c)) and abiotic conditions do not change rapidly across space. Its homogeneous, unfragmented surroundings makes it easier for individuals, like seed dispersers [80], to travel and contribute to the recovery of damaged habitat areas [81]. The ecosystem's  . Partial dependence plots of select attributes in table S1 that span the four attribute groups of highest importance in predicting the resistance and/or stability of the 5 × 5 km protected area ecosystems of interest across Africa ( figure 3; table S2). Each plot shows how ecosystem resistance or stability values change across the values of an attribute, after accounting for all other attributes in a trained regression model. Trends in each plot are depicted with a LOESS regression curve. The trends with respect to resistance are largely mirror images of those with respect to stability across all attributes, both here and at 1 × 1 km resolution (figure S9). mammalian community is low in functional dispersion (figures 4(d) and S9(d)), which entails greater functional redundancy [73]. Thus, ecological roles that foster ecosystem recovery, like seed dispersal, are shared by more species in the community, making it less likely that a disturbance could eliminate any such role and its effect on recovery entirely [82]. These stability-inducing features characterize protected area ecosystems occurring in regions like the east coast of Madagascar (figures 1(b) and S1(b)), including those located in Andasibe-Mantadia National Park or the Ambositra-Vondrozo Forest Corridor.
Identifying ecosystems like these that occur at the more extreme ends of the resistance-stability tradeoff is key if we are to conserve Africa's protected area environments. Ecosystems that are highly resistant are at risk of facing future disturbances before having fully recovered from prior ones, while those that are stable are at risk of substantial interferences to their function. Avoiding permanent losses of function [17,19,20] could be achieved if we promote characteristics of stability in resistant ecosystems, and vice versa, in a balanced manner. To illustrate, consider the ecosystems in Kruger National Park, which are increasingly being exposed to severe drought [45]. While highly resistant, their rapid recovery is possible as well if conservation initiatives foster the spread and growth of subdominant, yet drought-tolerant, grass species [45]. Such initiatives, however, would need to consider that too much spread could produce means for invasive species to emerge and hamper resistance. Ultimately, strategies that carefully balance resistance and stability are important if Africa's protected areas are to maintain their unique and essential biodiversity [34,35].
Striking the resistance-stability balance is a challenge for conservation, but insights from arenas like urban sustainability suggest that it is not insurmountable. An emerging framework in urban planning compares systems that are 'fail-safe' versus 'safe-tofail.' 'Fail-safe' systems focus on resistance, as they aim to remain unmoved by disturbance up-front [83,84]. They are characterized by armored and rigid structures, but if they fail, the failure becomes catastrophic. The levees surrounding New Orleans during Hurricane Katrina were built in a 'fail-safe' manner: while their sturdy design made them powerful, it also triggered their erosion, inability to adapt to poor foundational soil, and eventual breach [85]. 'Safeto-fail' systems are stable, as they focus on recovery and adaptation following failure [83,84,86]. Their structures are generally more flexible, making failures less catastrophic but more frequent [84]. An ideal system would integrate both 'fail-safe' and 'safe-to-fail' components, to minimize failure's frequency and impact. Pagodas in Japan, for example, contain both rigid base stones and flexible wooden joints [87]. In the face of seismic events, this ensures that pagodas are sturdy enough to avoid collapse and malleable enough to sway and rebound undamaged. Africa's ecosystems can similarly be assembled to comprise both resistant and stability-inducing elements. For instance, ecological corridors [88], which promote stability by facilitating movement of select species across space, can be integrated sparingly into resistant ecosystems in regions of high habitat heterogeneity [69].

Conclusions
In conclusion, our study demonstrates that resistance and stability are inherently opposing properties of ecosystems. We establish that the theory of a resistance-stability tradeoff is a widespread phenomenon across diverse ecosystems and ecosystem characteristics. Consequently, ecosystems rely on a balance between resistance and stability. Overemphasis of one suppresses the other, which in turn could engender permanent loss of ecosystem function. Finding the resistance-stability balance is critical in Africa's protected area ecosystems, which represent areas of global importance for biodiversity conservation. Time is of the essence, as anthropogenic disturbances to natural systems are intensifying and becoming more unpredictable. If ecosystems are to overcome disturbances and continue performing their functions, we must acknowledge that while resistance and stability oppose each other, both are essential and both must be upheld moving forward.

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: https://doi.org/10.5061/dryad.wh70rxwt4 [89]. The R code written to conduct all analyses, and the files on which that code depends, are available at the following URL: https://github.com/lauerd/ EcosystemResistanceStability [90]. resources, supervision, validation, writing-review & editing.